Six Markers Of AI Slop, And The Prompt Layer That Removes Them

A client can not always say what is wrong with a draft. They just stop replying. Then the retainer quietly ends, and the reason given is “we are taking content back in-house.”

Most of the time the writing was not slow, inaccurate or off-brief. It was slop: the flat, agreeable, faintly corporate dialect that a language model produces by default. Slop is hard to complain about precisely because nothing in it is technically wrong. It is just obviously not the person whose name is on it.

The useful thing about slop is that it is not a vague quality problem. It is a short list of recurring, nameable patterns — which means it can be constrained away rather than edited away.

The six markers

1. Throat-clearing openings. The first two sentences state something nobody disputes before the paragraph begins. “Content creation is vital for modern businesses. In this post, we will explore…” Delete until the first real claim; that is usually your actual opening line.

2. Symmetrical cadence. Every sentence lands between fourteen and eighteen words. No fragments, no long build, no short punch. Human writing is rhythmically uneven, and the evenness is what reads as synthetic even when a reader can not name it.

3. Metaphor reflex. Navigating turbulent waters. A double-edged sword. Unlocking potential. A beacon of hope. These arrive automatically because they are statistically safe, and they carry no information.

4. Hollow transitions. Furthermore. Moreover. It is worth noting that. Delving deeper. Each one occupies the position of an argument without making one. Cut them and check whether the logic still holds — usually it does, which proves they were never load-bearing.

5. Enthusiastic sycophancy. Thrilling. Groundbreaking. A testament to. Adjectives that signal excitement in place of conviction. A practitioner writes “this cut our review time in half”; a model writes “this is a truly exciting development.”

6. False hedging. “While some might disagree, it could be argued that…” The hedge is applied to claims that were not controversial in the first place. Real hedging is specific and rare; reflexive hedging drains every sentence equally.

Why “write in a bold, concise voice” does not work

Because bold and concise are subjective. The model has no shared referent for them, so it produces its own average of what those words usually accompany — which is the same default register you were trying to escape. Adjectives describe an outcome; they do not specify a behaviour.

Two things do specify behaviour: examples and prohibitions.

Replace adjectives with a negative-constraint block

A negative-constraint block is a fixed list of banned patterns that you paste into every request, rather than re-explaining your preferences each time. Its value is that it is boring and identical on every run — the output stops drifting between sessions.

CONSTRAINTS — apply to every draft:

DO NOT open with context-setting or a statement of what
   the piece will cover. Begin at the first real claim.
DO NOT use: furthermore, moreover, delving, it is worth
   noting, in today's landscape, game changer, unlock,
   navigate, testament to, thrilling, groundbreaking.
DO NOT use metaphors unless the metaphor is drawn from
   the specific industry being discussed.
DO NOT hedge a claim unless the uncertainty is real and
   you can name its source.
DO vary sentence length deliberately: at least one
   sentence under 8 words per paragraph.
DO prefer concrete nouns and plain verbs: build, run,
   cut, sell, ship. Not: utilise, leverage, facilitate.
DO write from the position of someone who has done this
   work, not someone describing it from outside.

Prohibitions outperform instructions here because they are checkable. “Be concise” can not be verified; “the word moreover does not appear” can, by you or by the model on a second pass.

Then give it three samples, not three adjectives

If you are writing as someone else, collect three pieces of their existing output: one where they explain their worldview (a talk, an interview, a manifesto), one where they walk through a process step by step, and three to five of their short social posts. Ask the model to extract the observable patterns — average sentence length and variance, recurring phrases, how they open and close, whether they use fragments or em dashes — into a written profile. Then keep that profile and reuse it.

The difference is that a profile is descriptive and reusable. An adjective is a wish repeated at the start of every session.

The pass you should not automate

Constraints and a voice profile get a draft to defensible. They do not get it to publishable, and treating them as if they do is how volume quietly destroys a retainer.

Budget about fifteen minutes per piece for a human pass that does three things a model can not: cut the paragraph that is competent but adds nothing, insert the one specific detail only you or the client know, and check that every factual claim is one you would defend in a meeting. That last one matters most — a fluent, confident, wrong sentence is far more damaging than an awkward one.

Where to start

Take a draft you already have. Run the six markers against it as a checklist and count the hits. Most first drafts return four or five. Then paste the constraint block above and regenerate the same piece. The gap between those two versions is the whole argument, and it takes about ten minutes to see for yourself.